Predicting COVID-19 Disease Progression and Patient Outcomes based on Temporal Deep Learning

doi:10.21203/rs.3.rs-44308/v2

Download PDF

Technical advance

Predicting COVID-19 Disease Progression and Patient Outcomes based on Temporal Deep Learning

https://doi.org/10.21203/rs.3.rs-44308/v2

This work is licensed under a CC BY 4.0 License

Journal Publication

published 08 Feb, 2021

Read the published version in BMC Medical Informatics and Decision Making →

You are reading this older preprint version

Read the latest preprint version →

Background: The coronavirus disease 2019 (COVID-19) pandemic has caused health concerns worldwide since December 2019. From the beginning of infection, patients will progress through different symptom stages, such as fever, dyspnea or even death. Identifying disease progression and predicting patient outcome at an early stage helps target treatment and resource allocation. However, there is no clear COVID-19 stage definition, and few studies have addressed characterizing COVID-19 progression, making the need for this study evident.

Methods: We proposed a temporal deep learning method, based on a time-aware long short-term memory (T-LSTM) neural network and used an online open dataset, including blood samples of 485 patients from Wuhan, China, to train the model. Our method can grasp the dynamic relations in irregularly sampled time series, which is ignored by existing works. Specifically, our method predicted the outcome of COVID-19 patients by considering both the biomarkers and the irregular time intervals. Then, we used the patient representations, extracted from T-LSTM units, to subtype the patient stages and describe the disease progression of COVID-19.

Results: Using our method, the accuracy of the outcome of prediction results was more than 90% at 12 days and 98%, 95% and 93% at 3, 6, and 9 days, respectively. Most importantly, we found 4 stages of COVID-19 progression with different patient statuses and mortality risks. We ranked 40 biomarkers related to disease and gave the reference values of them for each stage. Top 5 is Lymph, LDH, hs-CRP, Indirect Bilirubin, Creatinine. Besides, we have found 3 complications - myocardial injury, liver function injury and renal function injury. Predicting which of the 4 stages the patient is currently in can help doctors better assess and cure the patient.

Conclusions: To combat the COVID-19 epidemic, this paper aims to help clinicians better assess and treat infected patients, provide relevant researchers with potential disease progression patterns, and enable more effective use of medical resources. Our method predicted patient outcomes with high accuracy and identified a four-stage disease progression. We hope that the obtained results and patterns will aid in fighting the disease.

Medical Informatics

COVID-19

Disease progression

Outcome early prediction

Irregularly sampled time series

Time-aware long short-term memory

Due to technical limitations, full-text HTML conversion of this manuscript could not be completed. However, the latest manuscript can be downloaded and

accessed as a PDF.

Table 1 The conclusion of machine learning methods used in COVID-19 prediction tasks¹

Non-deep learning methods

Deep learning methods

Statistics

Regression

SVM

Decision tree

Basic NN

CNNs

RNNs

LDA [38]

NB [40]

k-NN [42]

LR [36,37]

Cox [32]

[36]

RF [36]

XGBoost [1]

BPNN [28,32,36],

GRNN [29],

RBFNN [29],

PNN [29]

Basic CNN [30],

Transfer CNN [31],

GDCNN [33],

COVID-Net [34],

COVNet [35]

¹ We use the abbreviations of methods and the full names are listed in Table 1.

Table 2 Demographic, laboratory and outcome information of 375 samples in training dataset

Characteristics^*	Statistics⁺	Characteristics^*	Statistics⁺
Demographics		Outcomes
Age, mean (years) Gender	58.83 Male 224; Female 151	Survival, count and rate Mortality, count and rate	201, 53.6% 174, 46.4%
Lab test mean (min, max) patient’s last measurements
cTnI Hemoglobin Serum chloride Prothrombin time procalcitonin Eosinophils(%) sIL－2R Alkaline phosphatase albumin basophil(%) Interleukin 10 Total bilirubin Platelet count monocytes(%) antithrombin Interleukin 8 indirect bilirubin RDW Neutrophils(%) Total protein Anti-TP Prothrombin activity HBsAg mean corpuscular volume hematocrit WBC Tumor necrosis factorα MCHC fibrinogen Interleukin 1β Urea lymphocyte count PH value Red blood cell count Eosinophil count Corrected calcium Serum potassium	747.76 (1.9, 43905.19) 124.89 (6.4, 178.0) 102.73 (77.7, 138.2) 16.01 (11.69, 84.22) 0.99 (0.02, 49.34) 0.56 (0.0, 5.61) 961.52 (61.0, 5608.04) 84.86 (37.18, 481.5) 33.06 (19.1, 45.07) 0.21 (0.0, 1.38) 12.5 (5.0, 500.0) 16.28 (2.95, 276.0) 187.78 (1.2, 472.5) 6.66 (0.62, 31.62) 87.57 (20.0, 136.0) 88.08 (5.0, 6385.85) 6.85 (1.0, 79.3) 12.94 (10.91, 22.91) 76.09 (15.9, 98.1) 66.06 (36.7, 79.33) 0.17 (0.02, 8.74) 80.97 (16.3, 136.64) 10.35 (0.0, 250.0) 89.75 (62.3, 110.73) 37.0 (19.9, 51.3) 15.18 (0.8, 1726.6) 12.07 (4.0, 168.0) 344.17 (286.0, 464.33) 4.38 (0.5, 8.89) 6.56 (5.0, 79.44) 9.74 (2.17, 68.4) 1.06 (0.13, 35.53) 6.47 (5.0, 7.54) 7.58 (0.1, 749.5) 0.03 (0.0, 0.33) 2.34 (1.78, 2.67) 4.45 (3.1, 9.86)	glucose neutrophils count Direct bilirubin Mean platelet volume ferritin RBCW-SD Thrombin time Lymphocyte(%) Anti-HCV D-D dimer Total cholesterol AST Uric acid HCO3- calcium NT-proBNP LDH platelet large cell ratio Interleukin 6 FDP monocytes count PLT distribution width globulin γ-GT INR basophil count(#) 2019-nCoV nucleic acid MCH APTT hs-CRP anti-HIV serum sodium thrombocytocrit ESR GPT eGFR creatinine	8.37 (2.43, 32.37) 7.03 (0.85, 31.43) 9.5 (1.7, 216.3) 10.89 (9.04, 14.0) 1634.37 (26.8, 42402.91) 41.78 (32.5, 83.3) 18.17 (13.0, 133.67) 16.45 (0.6, 54.83) 0.13 (0.03, 1.85) 6.2 (0.21, 29.82) 3.66 (0.66, 6.11) 56.53 (8.0, 1858.0) 297.47 (57.0, 1001.0) 22.59 (6.3, 29.7) 2.1 (1.35, 2.5) 3166.26 (5.57, 70000.0) 492.12 (143.0, 1867.0) 31.63 (16.48, 54.1) 130.37 (1.5, 5000.0) 44.29 (4.0, 182.4) 0.54 (0.08, 22.75) 12.97 (8.42, 25.3) 32.96 (13.8, 46.04) 51.79 (10.6, 555.25) 1.3 (0.86, 10.51) 0.02 (0.0, 0.12) -1.0 (-1.0, -1.0) 30.9 (20.8, 47.67) 41.72 (25.59, 100.27) 72.49 (0.2, 320.0) 0.1 (0.05, 0.26) 141.06 (122.8, 171.4) 0.21 (0.05, 0.49) 32.33 (2.0, 102.0) 42.6 (5.0, 1508.0) 81.07 (2.0, 164.7) 120.38 (39.25, 1497.0)

^* Characteristics have three types - demographics (age and gender), outcomes (survival and mortality) and laboratory test (74 items).

⁺ Statistics is statistical data for corresponding characteristics, such as rate, mean value and range. The statistical methods are described in each column.

Table 3 AUC-ROC of COVID-19 mortality prediction results by using baselines

	0 days early^*	3 days early^*	6 days early	9 days early	12 days early
Cox¹	0.9550.06	0.9920.02	0.870.01	0.85.01	0.810
k-NN¹	0.9500.02	0.9090.01	0.890.02	0.840.02	0.816
SVM¹	0.9690.04	0.9540.02	0.9300.03	0.895.04	0.857
DT¹	0.974	0.9590.03	0.924.00	0.897.01	0.869
BPNN¹	0.980.02	0.954.05	0.9331	0.894	0.878
PNN¹	0.985.01	0.961.02	0.9402	0.889	0.889
RNN¹	0.985	0.960.01	0.931.00	0.910	0.871
LSTM¹	0.990.01	0.961	0.937.02	0.920	0.897
T-LSTM²	0.9970.00	0.9690.01	0.947	0.921	0.914

^*n days early: The models make prediction n days before the final death/survival time. They use sequence data from day 0 to n days before the last time to predict.

¹ Cox: Cox's proportional hazards regression model is semi parametric regression model. It can analyze the influence of many factors on outcomes. It is used in [32].

¹ k-NN: k-Nearest Neighbors method makes prediction based on the information of nearest k samples in training set. In this mortality prediction task, the most accurate results appeared when k = 3.

¹ SVM: Support Vector Machines classify by solving the separation hyperplane which can divide the training data correctly and has the largest geometric interval.

¹ DT: Decision tree is a simple classifier consisting of sequences of hierarchically organized binary decisions. It is used in [6].

¹ BPNN: Back Propagation Neuron Network makes the signal and the error propagate forward and backward separately. It is used in [28].

¹ PNN: Probabilistic Neural Network is a forward propagation network and does not need back propagation to optimize parameters by using Bayesian decision-making. It is used in [29].

¹ RNN: Recurrent Neural Network have been introduced in the ‘T-LSTM’ section.

¹ LSTM: Long Short-Term Memory which we have introduced in the ‘T-LSTM’ section. Here, the hyperparameter setting is same as T-LSTM.

²T-LSTM: Time-aware LSTM is the model used in this paper. Its inputs are the three-dimensional vectors and the time intervals. The values for each dimension are the values of LDH, lymphocyte and hs-CRP in patients’ blood tests. Its output is the binary result 0/1. Here, 0 indicates survival and 1 indicates death. The hidden states in its units are 64 dimensional, and the fully connected layer has 32 dimensions.

Table 4 AUC-ROC of COVID-19 mortality prediction results by using T-LSTM on different sets at different timestamps

	Training⁺	Validation⁺	Test⁺
0 days early^*	0.9960.01	0.9970.01	0.9970.00
3 days early	0.9890.00	0.9870.01	0.9690.01
6 days early	0.960	0.9572	0.947
9 days early	0.944	0.9350.01	0.921
12 days early	0.9260.01	0.9242	0.914
15 days early	0.8910.01	0.8830.01	0.863
18 days early	0.8520.01	0.8342	0.8192

^*n days early: The model makes a prediction n days before the final death/survival time. It uses sequence data from day 0 to n days before the last time to predict.

⁺ We use the records of 375 patients as the training set; the ratio of training set to verification set is 0.8:0.2. The records of 110 patients make up the test set. This experiment is conducted on 5-fold cross-validation.

Table 5 AUC-ROC of COVID-19 mortality prediction results by using T-LSTM with 40 or 3 laboratory tests⁺

	0 days early^*	3 days early^*	6 days early	9 days early	12 days early
40 features	0.9490.01	0.9200.03	0.9150.01	0.9100.01	0.9030.01
3 features	0.9970.00	0.9690.01	0.947	0.921	0.914

^*n days early: The model makes prediction n days before the final death/survival time. It uses sequence data from day 0 to n days before the last time to predict.

⁺ The inputs of T-LSTM are time series of 48 laboratory tests or 3 laboratory tests (LDH, lymph and hs-CRP).

Table 6 Feature statistics of patients in different stages of COVID-19 disease progression

Survival class (general)
Mean value	Stage 1	Stage 2	Stage 3	Stage 4
Mortality rate (%)	7.57	0	1.91	0
Time distance (days)	22.65	14.08	7.98	2.75
Lymph (%)	16	18	21	31
LDH (U/l)	328	301	245	199
hs-CRP (mg/l)	43	39	21	3
Indirect Bilirubin (μmol/l)	7	6	4	3
Creatinine (μmoI/l)	98	75	89	76
INR	2	1	1	1
Serum Sodium (mmol/l)	138	139	140	137
eGFR (ml/min)	79	109	112	111
Serum Chlorine (mmol/l)	97	99	103	102
Albumin (g/l)	41	39	38	40
Death class (critical)
Mean value	Stage 1	Stage 2	Stage 3	Stage 4
Mortality rate (%)	76.32	88.76	91.28	100
Time distance (days)	26.96	18.76	9.32	2.05
Lymph (%)	15	10	9	4
LDH (U/l)	338	364	375	499
hs-CRP (mg/l)	48	55	69	84
Indirect Bilirubin (μmol/l)	8	9	14	23
Creatinine (μmoI/l)	104	106	120	125
INR	2	2	3	2
Serum Sodium (mmol/l)	140	140	135	129
eGFR (ml/min)	75	71	70	57
Serum Chloride (mmol/l)	96	103	104	105
Albumin (g/l)	40	32	33	32

Table 7 Ranking of average KL divergence values of top 40 features

Ranking	Feature	Average KL	Ranking	Feature	Average KL
1	Lymph	0.0421	21	Total Cholesterol	0.0041
2	LDH	0.0392	22	Interleukin 6	0.0032
3	hs-CRP	0.0376	23	I sIL－2R	0.0031
4	Indirect Bilirubin	0.0324	24	cTnI	0.0030
5	Creatinine	0.0302	25	RBCW-SD	0.0024
6	INR	0.0235	26	Uric Acid	0.0023
7	Serum Sodium	0.0232	27	Corrected Calcium	0.0022
8	eGFR	0.0225	28	Interleukin 8	0.0019
9	Serum Chloride	0.0224	29	Prothrombin Time	0.0018
10	Albumin	0.0193	30	Serum Potassium	0.0017
11	Globulin	0.0177	31	Interleukin 1β	0.0017
12	Hematocrit	0.0122	32	D-D dimer	0.0016
13	Hemoglobin	0.0091	33	FDP	0.0016
14	Fibrinogen	0.0079	34	Antithrombin	0.0015
15	γ-GT	0.0079	35	Procalcitonin	0.0010
16	ESR	0.0078	36	Platelet Count	0.0009
17	NT-proBNP	0.0074	37	WBC	0.0006
18	APTT	0.0053	38	Ferritin	0.0005
19	Eosinophils	0.0051	39	Interleukin 10	0.0004
20	basophil	0.0049	40	PLT	0.0004